AI compute has continued to scale by orders of magnitude, delivering better frontier models. Training clusters now span hundreds of thousands of GPUs and desire to reach 1 million+. Scale-up domains want to grow from 72 GPUs to larger, higher-bandwidth domains with a single tier of switches. If these levels of scale can be achieved, new MoE models with trillions of parameters, more experts, and more active experts per token are possible – delivering greater intelligence. Yet switch silicon has advanced only incrementally, doubling radix and capacity every 2–2.5 years, creating a “network wall” stifling these ambitions.

We will introduce Eridu and a new class of switch delivering a step function increase in radix and throughput to breakthrough this network wall and unlock the next era of AI infrastructure. We will discuss how a high-radix switch can enable larger, more capable AI models while improving GPU utilization, job completion time, power efficiency, and cluster economics.